README.md

---
license: apache-2.0
language:
- en
pretty_name: OmitI2V
size_categories:
- n<1K
task_categories:
- image-to-video
tags:
- text-to-video
- image-to-video
- text-guided-image-to-video
- benchmark
- prompt-adherence
- semantic-fidelity
- video-generation
---

# OmitI2V

**OmitI2V** is a benchmark for evaluating **semantic fidelity** (prompt adherence) in **text-guided image-to-video (TI2V)** generation, focusing on prompts that require *substantial* edits to the reference image — object **addition**, **deletion**, and **modification**. It is the benchmark introduced in *AlignVid: Taming Visual Dominance via Training-Free Attention Modulation in Text-guided Image-to-Video Generation* (**ICML 2026**).

| | |
|---|---|
| 📄 Paper | <https://arxiv.org/abs/2512.01334> |
| 💻 Code | <https://github.com/LAW1223/AlignVid> |
| 🌐 Project page | <https://law1223.github.io/AlignVid/> |

It contains **367 human-annotated samples**. Each sample pairs a reference image with an editing instruction and a set of yes/no VQA questions that probe whether the requested edit is actually realized in the generated video.

## Dataset structure

```
.
├── meta.json        # 367 annotated samples
├── modification/    # reference images, organized by sub-category / domain
├── addition/
└── deletion/
```

Images are referenced by the `image-path` field in `meta.json`, **relative to the dataset root** (e.g., `modification/pose/human/sample_0.jpg`).

## Fields (per entry in `meta.json`)

| field | type | description |
|---|---|---|
| `id` | str | unique sample id (e.g., `sample_0`) |
| `image-path` | str | reference image path, relative to the dataset root |
| `prompt` | str | the text instruction driving the video |
| `main-category` | str | one of `modification`, `addition`, `deletion` |
| `sub-category` | str | finer edit type (e.g., `pose`, `appearance`, `element`) |
| `domain` | str | content domain (e.g., `human`, `animal`, `nature`, `building`) |
| `type` | str | image source: `real image`, `generated image`, or `animation image` |
| `key` | list[str] | short keyword(s) summarizing the target change |
| `expected-change` | str | natural-language description of the expected edit |
| `resolution` | str | image resolution |
| `aspect_ratio` | str | image aspect ratio |
| `questions` | list[dict] | VQA items, each `{id, question, expected_answer, category}` |

## Statistics

- **Samples:** 367
- **Main category:** modification 113 · addition 129 · deletion 125
- **Image type:** real image 290 · animation image 56 · generated image 21
- **Domains:** human, animal, nature, object, building, animation, environment, effect, and more.

## Usage

```python
import json

meta = json.load(open("meta.json", encoding="utf-8"))
print(len(meta))                 # 367
ex = meta[0]
print(ex["prompt"])              # editing instruction
print(ex["image-path"])          # e.g. modification/pose/human/sample_0.jpg
for q in ex["questions"]:
    print(q["question"], "->", q["expected_answer"])
```

The `questions` provide a yes/no protocol for measuring fine-grained edit compliance of a generated video against the prompt.

## Evaluation

The full evaluation pipeline (VQA-based semantic fidelity with Qwen2.5-VL, plus VBench-style visual-quality metrics) and ready-to-edit inference scripts for FramePack, FramePack-F1, and Wan2.1 live in the AlignVid repository:

<https://github.com/LAW1223/AlignVid>

## License

Released under the Apache-2.0 License.

## Citation

```bibtex
@article{liu2025alignvid,
  title={AlignVid: Training-Free Attention Scaling for Semantic Fidelity in Text-Guided Image-to-Video Generation},
  author={Liu, Yexin and Shu, Wen-Jie and Huang, Zile and Zheng, Haoze and Wang, Yueze and Zhang, Manyuan and Lim, Ser-Nam and Yang, Harry},
  journal={arXiv preprint arXiv:2512.01334},
  year={2025}
}
```